Intention-to-treat analysis

An intention-to-treat analysis compares participants according to their randomized assignment, retaining that assignment even when they do not use the assigned intervention. Its causal effect is the effect of assignment under the study's actual adherence pattern, rather than the effect of perfect adherence. Randomization protects this comparison from baseline confounding; missing data can still require assumptions or additional analysis.

Treatment contamination occurs when participants in one arm of a randomized controlled trial receive an intervention assigned to another arm. This is distinct from contamination in robust estimation. The intention-to-treat analysis still compares randomized assignments, but intervention receipt becomes less separated between the arms, often diluting the assignment effect. Receipt-based comparisons can introduce confounding because receipt need not be randomized.

Assume treatment receipt gives event risk $p_1$, nonreceipt gives risk $p_0$, and a fraction $c$ in each equally sized randomized arm crosses to the opposite treatment. The two assignment risks are $(1-c)p_1+c{p}_0$ and $c{p}_1+(1-c)p_0$. Subtracting yields the displayed attenuation, with ${\Delta }_{\mathrm{receipt}}=p_0-p_1$. It depends on the stated equal crossover fractions and unchanged receipt-specific risks; it is not a universal formula for every form of treatment contamination.

A treatment consequence is not interchangeable with a baseline covariate in an intention-to-treat analysis. Let randomized $Z$ be independent of $U,\epsilon$, and let actual intervention use be $D=Z+U$, with $Y=D+\epsilon$ and independent centered errors. The assignment changes the mean outcome by one. But the conditional expectation $\mathbb{E}[Y\mid Z,D]=D$ has coefficient zero on $Z$. Thus adjustment for the post-randomization variable $D$ changes the estimand even in a simple identified linear regression.